How to join on closest date in Postgresql - postgresql

Question

Suppose, I have following tables
product_prices
product|price|date
-------+-----+----------
apple |10 |2014-03-01
-------+-----+----------
apple |20 |2014-05-02
-------+-----+----------
egg |2 |2014-03-03
-------+-----+----------
egg |4 |2015-10-12
purchases:
user|product|date
----+-------+----------
John|apple |2014-03-02
----+-------+----------
John|apple |2014-06-03
----+-------+----------
John|egg |2014-08-13
----+-------+----------
John|egg |2016-08-13
What I need is table similar to this:
name|product|purchase date |price date|price
----+-------+--------------+----------+-----
John|apple |2014-03-02 |2014-03-01|10
----+-------+--------------+----------+-----
John|apple |2014-06-03 |2014-05-02|20
----+-------+--------------+----------+-----
John|egg |2014-08-13 |2014-08-13|2
----+-------+--------------+----------+-----
John|egg |2016-08-13 |2015-10-12|4
Or "what is the price for product at this day". Where price is calculated based on date from products table.
On real DB I tried to use something similar to:
SELECT name, product, pu.date, pp.date, pp.price
FROM purchases AS pu
LEFT JOIN product_prices AS pp
ON pu.date = (
SELECT date
FROM product_prices
ORDER BY date DESC LIMIT 1);
But I keep either getting only left part of table (with (null) instead of product dates and prices) or many rows with all the combinations of prices and dates.

Answer 1

I would suggest changing product_prices table to use a daterange column instead (or at least a start_date and an end_date).
You can use an exclusion constraint to make sure you never have overlapping ranges for one product and an insert trigger that "closes" the "current" prices and creates a new unbounded range for the newly inserted price.
A daterange can efficiently be indexed and with that in place the query gets as easy as:
SELECT name, product, pu.date, pp.valid_during, pp.price
FROM purchases AS pu
LEFT JOIN product_prices AS pp ON pu.date <# pp.valid_during
(assuming the range column is named valid_during)
The exclusion constraint would only work however if the product was an integer (not a varchar) - but I guess your real product_purchases table uses a foreign key to some product table anyway (which is an integer).
The new table definitions could look something like this:
create table purchase_prices
(
product_id integer not null references products,
price numeric(16,4) not null,
valid_during daterange not null
);
And the constraint that prevents overlapping ranges:
alter table purchase_prices
add constraint check_price_range
exclude using gist (product_id with =, valid_during with &&);
The constraint needs the btree_gist extension.
As always improving query speed comes with a price and in this case it's the higher maintenance costs for the GiST index. You would need to run some tests to see if the easier (and most probably much faster) query outweighs the slower insert performance on purchase_prices.

Answer 2

Look at your scalar sub-query very closely. It is not correlated back to the outer query. In other words, it will return the same result every time: the latest date in the product_prices table. Period. Think about the query out of context:
SELECT date
FROM product_prices
ORDER BY date DESC LIMIT 1
There are two problems with it:
It will return 2015-10-12 for every row in the join and ultimately, nothing was purchased on that date, hence, null.
Your approximation of closest is that the dates are equal. Unless you have a product_prices row for every product for every single date, you'll always have misses. "Closest" implies distance and ranking.
WITH close_prices_by_purchase AS (
SELECT
p.user,
p.product,
p.date pp.date,
pp.price,
row_number() over (partition by pp.product, order by pp.date desc) as distance -- calculate distance between purchase date and price date
FROM purchases AS p
INNER JOIN product_prices AS pp on pp.product = p.product
WHERE pp.date < p.date
)
SELECT user as name, product, pu.date as purchase_date, pp.date as price_date, price
FROM close_prices_by_purchase AS cpbp
WHERE distance = 1; -- shortest distance

Answer 3

You can try something like this, although I am sure there's a better way:
with diffs as (
select
a.*,
b."date" as bdate,
b.price,
b."date" - a."date" as diffdays,
row_number() over (
partition by "user", a."product", a."date"
order by "user", a."product", a."date", b."date" - a."date" desc
) as sr
from purchases a
inner join product_prices b on a.product = b.product
where b."date" - a."date" < 1
)
select
"user" as "name",
product,
"date" as "purchase date",
bdate as "price date",
price
from diffs
where sr = 1
Example: https://www.db-fiddle.com/f/dwQ9EXmp1SdpNpxyV1wc6M/0
Explanation
I attempted to join both tables and find the difference between dates of purchase and price, and ranked them by closest date prior to the purchase. Rank of 1 will go to the closest date. Then, data with rank of 1 was extracted.

Answer 4

This is a great place to use date ranges! We know the start date of the price range and we can use a window function to get the next date. At that point, it's really easy to figure out the price on any day.
with price_ranges as
(select product,
price,
date as price_date,
daterange(date, lead(date, 1)
OVER (partition by product order by date), '[)'
) as valid_price_range from product_prices
)
select "user" as name,
purchases.product,
purchases.date,
price_date,
price
from purchases
join price_ranges on purchases.product = price_ranges.product
and purchases.date <# price_ranges.valid_price_range
order by purchases.date;